Microsoft has unveiled Maia 300, the third generation of its proprietary artificial intelligence chips. The announcement marks another step in the company's strategy to reduce its dependence on Nvidia and AMD processors in its data centers.
The launch is significant at a time when major tech companies are competing to control every piece of the AI supply chain, from models to the hardware that runs them. In recent months, Microsoft, OpenAI, Google, and Amazon have intensified their efforts to develop their own silicon to avoid being tied to a single external provider.
Up to 40% more efficiency than leading GPUs
According to Scott Guthrie, executive vice president of Microsoft's Cloud and AI group, the company expects Maia 300 to offer 40% more tokens per megawatt compared to leading GPU-based systems. This is an internal company estimate, not an independent measurement verified by third parties.
Guthrie also specified that, in a full 96-accelerator configuration, the new system can achieve up to seven times the rack-scale token performance compared to Maia 200, its previous generation.
The chip is designed for what Microsoft calls its primary AI inference system, which is the phase where an already trained model responds to user queries. This stage accounts for a significant portion of the daily operating costs of cloud services.
Production deployment starting in 2027
Maia 300 will not be available immediately. Microsoft expects the chip to begin being used in production workloads in 2027, so for now, this is a roadmap announcement.
Guthrie was clear about the scope of the plan: the company will scale its Azure Maia family toward multi-gigawatt deployments starting that year, with the goal of making it the primary AI inference system for its own products and services.
This does not mean that Microsoft will stop buying chips from Nvidia or AMD, but it does indicate that their relative weight within the company's infrastructure will progressively decrease.
Chiplet architecture on TSMC’s N2P process
The efficiency improvement promised by Microsoft is supported by the chip's architecture. Maia 300 is manufactured using TSMC's N2P process and uses a chiplet design that separates compute, memory controller, and input/output functions.
The Azure hardware and infrastructure team explains that this approach allows more silicon area to be dedicated to compute performance and ensures each function uses the most suitable technology for its role. In practice, this avoids wasting chip surface area on tasks that are not pure computation, combining different technologies as needed.
Maia 200 is already operating in Copilot and GitHub Copilot
Maia 300 arrives as the successor to Maia 200, which Microsoft introduced in January. That previous generation is already running production workloads in Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and the company's own AI models.
Guthrie described Maia 200 as the best AI infrastructure in Microsoft's fleet, measured by real AI service metrics, although this is an internal assessment by the company itself.
Regarding Maia 300, Microsoft defines it as a complete AI system designed for new types of workloads: long-context reasoning, latency-sensitive conversational experiences, and agentic tasks that require several successive steps.
Suleyman’s commitment to controlling the entire technology stack
The new chip is also part of Microsoft's work with its own AI models, with which it seeks to compete with Google DeepMind, OpenAI, and Anthropic.
Mustafa Suleyman, CEO of Microsoft AI, defended the need to control the entire technology chain, including silicon. As he explained, at the scale at which the company operates, it is necessary to have a proprietary platform, co-designed with its models, so running Microsoft AI models on the Maia chip has become a priority for his team.
The reasoning behind this strategy is vertical integration: if the model and the chip are designed together, both can be adapted to each other to gain efficiency. We will have to wait for production data to see how much this translates into real cost savings.
What changes for Microsoft users
The announcement of Maia 300 has concrete implications that are worth noting.
Those who already use Copilot, GitHub Copilot, or Foundry are currently working on Maia 200 without having explicitly chosen it. The hardware change happens within Microsoft's infrastructure and does not require any action from the user.
For those who purchase AI capacity on Azure, the multi-gigawatt deployment planned for 2027 indicates that the offering will increasingly depend on the Maia chip. It may be useful to ask the provider which workload runs on which type of hardware before comparing prices between services.
For businesses that use models with extensive reasoning or agents that execute multi-step tasks, these are precisely the scenarios for which Microsoft claims to have designed Maia 300. This is a promise from the company that can be verified once the chip is in production.
An illustrative example: a support team that uses an AI agent to resolve incidents in several steps generates many more tokens per query than a simple chat. In such scenarios, a 40% improvement in tokens per megawatt could translate into lower energy costs per resolved incident, provided the figure announced by Microsoft is confirmed in practice.
What it implies for the future
The announcement does not change anything immediately, as Maia 300 will not reach production until 2027. What it does make clear is Microsoft's strategic direction, which is committed to making its own silicon the main engine of its artificial intelligence services.
The practical consequence for customers is that the cost and availability of AI on Azure will increasingly depend on chips developed by Microsoft itself and less on those from Nvidia and AMD. It will be important to observe whether the promised efficiency effectively translates into lower prices once the chip is deployed in data centers.
Frequently Asked Questions
What is Maia 300?
It is the third generation of artificial intelligence chips designed by Microsoft, intended for the inference phase, that is, for already trained AI models to respond to user queries.
When will Maia 300 be available?
Microsoft expects it to begin being used in production workloads starting in 2027. For now, it is a roadmap announcement, with no immediate availability for customers.
Will Maia 300 completely replace Nvidia chips at Microsoft?
Not necessarily. Microsoft has not announced that it will stop buying chips from Nvidia or AMD, but it has indicated that their relative weight in the company's infrastructure will decrease as Maia becomes the primary inference system.
What advantages does Maia 300 promise over current GPUs?
According to Microsoft, the chip would offer 40% more tokens per megawatt compared to leading GPU-based systems, as well as up to seven times more performance than its predecessor, Maia 200, in full-rack configurations.
Which services already use Microsoft’s Maia chips?
Maia 200, the previous generation, is already operating in Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and the company's own AI models.
Does this announcement affect Copilot users today?
Not directly. The hardware change occurs in Microsoft's cloud infrastructure and does not require any action from the end user.
